3D fish animation with visual learning ability

We presented a visual behavior learning mechanism, to learn the realistic motion from real fishes and to automatically generate 3D animation. This method does not need to set markers or sensors on the objects; neither needs the pre-created motion database. All the learning data are obtained from the multi-view input videos. We use multiple cameras to record the motion of fishes, and track the deformable objects by using Template matching associated with Selective binary and Gaussian filtering regularized level set. After tracking, the skeletons are extracted by the Delaunay triangulation from the contour of creatures. We also proposed a line fitting method to combine the 2D skeletons of two views into the 3D skeleton. Furthermore, the proposed mechanism analyzes the path data and skeleton motions to create the learning data. Finally, one can simulate the continuous animation that not limited to the time length of input videos.